Building and Refining Rhetorical-Semantic Relation Models

Proceedings of NAACL HLT 2007, pages 428–435, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Building and Refining Rhetorical-Semantic Relation Models Sasha Blair-Goldensohn and Google, Inc. 76 Ninth Avenue New York, NY sasha@google.com Kathleen R. McKeown† and Owen C. Rambow‡ † Department of Computer Science ‡ Center for Computational Learning Systems Columbia University {kathy,rambow}@cs.columbia.edu Abstract We report results of experiments which build and refine models of rhetorical- semantic relations such as Cause and Con- trast. We adopt the approach of Marcu and Echihabi (2002), using a small set of patterns to build relation models, and ex- tend their work by refining the training and classification process using parame- ter optimization, topic segmentation and syntactic parsing. Using human-annotated and automatically-extracted test sets, we find that each of these techniques results in improved relation classification accuracy. 1 Introduction Relations such as Cause and Contrast, which we call rhetorical-semantic relations (RSRs), may be sig- naled in text by cue phrases like because or how- ever which join clauses or sentences and explicitly express the relation of constituents which they con- nect (Example 1). In other cases the relation may be implicitly expressed (2).1 Example 1 Because of the recent accounting scan- dals, there have been a spate of executive resigna- tions. Example 2 The administration was once again be- set by scandal. After several key resignations ... 1The authors would like to thank the four anonymous re- viewers for helpful comments. This work was supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR0011-06-C-0023. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of DARPA. The first author performed most of the research reported in this paper while at Columbia University. In this paper, we examine the problem of detect- ing such relations when they are not explicitly sig- naled. We draw on and extend the work of Marcu and Echihabi (2002). Our baseline model directly implements Marcu and Echihabi’s approach, opti- mizing a set of basic parameters such as smoothing weights, vocabulary size and stoplisting. We then focus on improving the quality of the automatically- mined training examples, using topic segmenta- tion and syntactic heuristics to filter out training instances which may be wholly or partially in- valid. We find that the parameter optimization and segmentation-based filtering techniques achieve sig- nificant improvements in classification performance. 2 Related Work Rhetorical and discourse theory has a long tradition in computational linguistics (Moore and Wiemer- Hastings, 2003). While there are a number of differ- ent relation taxonomies (Hobbs, 1979; McKeown, 1985; Mann and Thompson, 1988; Martin, 1992; Knott and Sanders, 1998), many researchers have found that, despite small differences, these theories have wide agreement in terms of the core phenom- ena for which they account (Hovy and Maier, 1993; Moser and Moore, 1996). Work on automatic detection of rhetorical and dis- course relations falls into two categories. Marcu and Echihabi (2002) use a pattern-based approach in mining …

Content truncated to 500 words